Matt's Homepage

I’ve been exploring

Fish Audio S2 Pro is one of (if not the) best text-to-speech solutions. Getting it installed locally and working, however, isn’t so straightforward on Windows 10. There are at least 2 different ways to get this working. One of which is to download/run

Method 0: Use the free online version

It’s not hard – but expect to be limited in usage. https://fish.audio/app/text-to-speech

Method 1: Fish S2 Pro Zero Docker

1. Go to the Huggingface Fish Audio S3 Pro project page.
2. Ensure you’re logged into Huggingface, and you should see the ‘Run Locally’ option Go up in the link
3. Ensure Docker is installed on the Windows desktop and WSL support is enabled in the Docker options.
4. Open a WSL session running Ubuntu 24.04 or similar.
5. Enter the docker command:

docker run -it -p 7860:7860 --platform=linux/amd64 --gpus all \
registry.hf.space/artificialguybr-fish-s2-pro-zero:latest python app.py

6. You’ll see the docker container download along with the models and start up:

(base) me@DESKTOP:/mnt/c/fish-audio-s2$ docker run -it -p 7860:7860 --platform=linux/amd64 --gpus all registry.hf.space/artificialguybr-fish-s2-pro-zero:latest python app.py
Cloning into 'fish-speech'…
remote: Enumerating objects: 6605, done.
remote: Counting objects: 100% (1088/1088), done.
remote: Compressing objects: 100% (292/292), done.
remote: Total 6605 (delta 905), reused 796 (delta 796), pack-reused 5517 (from 2)
Receiving objects: 100% (6605/6605), 28.21 MiB | 10.42 MiB/s, done.
Resolving deltas: 100% (4328/4328), done.
Warning: You are sending unauthenticated requests to the HF Hub. Please set a HF_TOKEN to enable higher rate limits and faster downloads.
Fetching 13 files: 100%|████████████████████████████████████████████████████████████████| 13/13 [02:19<00:00, 10.72s/it]Fetching 13 files: 62%|████████████████████████████████████████ | 8/13 [02:19<01:13, 14.78s/itYou are using a model of type fish_qwen3_omni to instantiate a model of type `. This may be expected if you are loading a checkpoint that shares a subset of the architecture (e.g., loading asam2_video checkpoint intoSam2Model), but is otherwise not supported and can yield errors. Please verify that the checkpoint is compatible with the model you are instantiating. Download complete: : 11.0GB [02:19, 79.0MB/s] 2026-07-03 18:59:16.787 | INFO | fish_speech.models.text2semantic.llama:from_pretrained:504 - Injected Semantic IDs into Config: 151678-155773 2026-07-03 18:59:16.787 | INFO | fish_speech.models.text2semantic.llama:from_pretrained:520 - Loading model from /home/user/.cache/huggingface/hub/models--fishaudio--s2-pro/snapshots/1de9996b6be38b745688de084d87a5633f714e4e, config: DualARModelArgs(model_type='dual_ar', vocab_size=155776, n_layer=36, n_head=32, dim=2560, intermediate_size=9728, n_local_heads=8, head_dim=128, rope_base=1000000, norm_eps=1e-06, max_seq_len=32768, dropout=0.0, tie_word_embeddings=True, attention_qkv_bias=False, attention_o_bias=False, attention_qk_norm=True, codebook_size=4096, num_codebooks=10, semantic_begin_id=151678, semantic_end_id=155773, use_gradient_checkpointing=True, initializer_range=0.01976423537605237, is_reward_model=False, scale_codebook_embeddings=True, audio_embed_dim=2560, n_fast_layer=4, fast_dim=2560, fast_n_head=32, fast_n_local_heads=8, fast_head_dim=128, fast_intermediate_size=9728, fast_attention_qkv_bias=False, fast_attention_qk_norm=False, fast_attention_o_bias=False, norm_fastlayer_input=True) 2026-07-03 18:59:46.228 | INFO | fish_speech.models.text2semantic.llama:from_pretrained:552 - Loading sharded safetensors weights 2026-07-03 18:59:46.717 | INFO | fish_speech.models.text2semantic.llama:from_pretrained:588 - Model weights loaded - Status: <All keys matched successfully> 2026-07-03 18:59:48.707 | INFO | fish_speech.models.text2semantic.inference:init_model:366 - Restored model from checkpoint 2026-07-03 18:59:48.708 | INFO | fish_speech.models.text2semantic.inference:init_model:371 - Using DualARTransformer/usr/local/lib/python3.10/site-packages/torch/nn/utils/weight_norm.py:144: FutureWarning:torch.nn.utils.weight_normis deprecated in favor oftorch.nn.utils.parametrizations.weight_norm`.
WeightNorm.apply(module, name, dim)
* Running on local URL: http://0.0.0.0:7860, with SSR ⚡ (experimental, to disable set ssr_mode=False in launch())
* To create a public link, set share=True in launch().

7. Open a browser to localhost:7860

Method 2: Build and run locally

1. Clone the github project: https://github.com/fishaudio/fish-speech
2. Open a WSL Ubuntu 24.04 installation.
3. Ensure you have nVidia support in WSL installed. Rebooting after this is often required.
4. Follow the installation/build instructions.
   • Run the conda setup steps
   • Run the UV steps for CPU or GPU depending on your install
   • Skip the docker part
5. WebUI On the left menu, select the ‘Inference’ from the list of items
   • Download the model weights with the hf command
     • You can test using the command line inference steps if you want to test it
   • Scroll down to the WebUI inference
   • Install Gradio if you want the older style (not so much recommended, but easier than Awesome WebUI)
   • Install the ‘Awesome WebUI’
   • Start the ‘Awesome WebUI’ using the python command
     • Open a browser on http://localhost:8888/ui
6. Server
   • Select the ‘Server’ item from the list of left-hand items
   • Run the python command to start the server locally.
   • Try out one of the api_client.py commands to test it out

Things you can do with the server:

• Create your own local cloned voice (.npy files) from sample wav files and transcribed text from inference.py.
• Check out the Text to Speech API Developer’s Guide
  • The Text to Speech API
• Check out the emotional cues you can add to the text

Other links:

• List of emotive cues you might give: https://fish.audio/blog/fish-audio-s2-fine-grained-ai-voice-control-at-the-word-level/
• Reddit article

After 18 months, graphics features like multi-frame generation are working really well, but only a handful of games support path tracing to use it. How few? 7 titles. Even worse, the Steam list of top 100 most anticipated titles – only 1 has path tracing that allows multi-frame generation.

And yet, the price of 5090’s continues to remain sky high and selling better than ever; but almost certainly because of the 32gb of RAM for AI usage. I know that’s why I got my 5090.

With the huge implosing of the game development world, the rise of faster and cheaper to develop indie titles, its no wonder top-end graphics is taking a back seat.

In a report published in February that brings us one step closer to the Terminator and War Games movies, when presented scenarios, the leading open AI models used nuclear weapons in 95% of war-game scenarios. None chose full surrender, even when losing.

Researchers warn AI use may escalate conflicts under pressure.

While this was a parody/joke – it’s actually pointing in the direction software is likely headed. You can tell some of the folks in the video are seeing it. What if you just type in the application you want and it’s generated for you? Sure this is flaky now, but given the advancements as well as putting guard rails on things to guide it – this is all very much a possible future. You don’t buy software anymore and SaaS companies don’t exist either – you make the software you want by asking for it and the system generates exactly the app you need.

Download it here: https://vibeos.sh

This is a really interesting set of DEV articles written by a software engineer. Initially a set of 8 articles, it’s expanded to 21 and counting. It does an excellent job of describing and analyzing exactly the conditions engineers are going through – from watching an AI agent do a week’s worth of work in 1 hour while they were in a meeting to seeing 27.5% of programmer jobs disappear in 2 years.

Definitely a good read.

Before Mythos got shut down, Anthropic’s Opus 4.7 saw a huge spike in complaints about the company’s Acceptable Use Policy (AUP) classifier causing customers trouble in getting legitimate work done.

First, grab a trustworthy Gemma 4 gguf model (unsloth is great). I have been fooling around with Q4, Q6, Q8 models of gemma-4-26B and gemma-4-31B models.

operty	E2B	E4B	31B Dense
Total Parameters	2.3B effective (5.1B with embeddings)	4.5B effective (8B with embeddings)	30.7B
Layers	35	42	60
Sliding Window	512 tokens	512 tokens	1024 tokens
Context Length	128K tokens	128K tokens	256K tokens
Vocabulary Size	262K	262K	262K
Supported Modalities	Text, Image, Audio	Text, Image, Audio	Text, Image
Vision Encoder Parameters	~150M	~150M	~550M
Audio Encoder Parameters	~300M	~300M	No Audio

Grab the latest version of llama.cpp and compile it with CUDA support for GPU usage (or CPU if you don’t have a CUDA enabled GPU).

Then set up your server command line and it seems some get about 600 Tok/s using 26B on a 5090, or maybe try out the turbo-quant variant of llama-cpp (I have not).

Then take off and make some nifty projects with it.